Re-identification of fish individuals of undulate skate via deep learning within a few-shot context

Individual re-identification is critical to track population changes in order to assess status, being particularly relevant in species with conservation concerns and difficult access like marine organisms. For this, we propose photo-identification via deep learning as a non-invasive technique to discriminate between individuals of the undulate skate (Raja undulata). Nevertheless, accruing enough training samples might be difficult to achieve in the case of underwater fish images. We develop a novel methodology based on a siamese neural network that incorporates statistical fundamentals as motivation to overcome the few-shot context. Our work provides a hands-on experience and highlights on pitfalls when trying to apply photo-identification in a limited scenario, concerning both data quantity and quality, yet providing remarkable results over the test set including recaptures, where the model is capable of correctly identifying the 70% of the individuals. The findings of this study can be of strong impact for the research teams becoming familiar with deep learning approaches, as it can be easily extended to re-identify individuals of other marine species of interest from a conservation or exploitation point of view. •Deep Learning represents a non-invasive alternative to monitor populations.•Marine ecosystems often represent a data-limited scenario or a few-shot context.•Siamese networks establish the similarity between pairs of input images.•The statistical motivation behind ensemble methods can help deal with variability.•Proposed method is robust, minimising error and bias of animal photo-identification.